Who Builds the Honest Chatbot?
A subsidy auction game about AI chatbots that flatter their users.
The story
Many AI chatbots are sycophantic: they agree with and flatter users because agreeable answers keep people chatting, which wins market share. A government wants one company to build an honest chatbot instead. It has one grant of at most points, so it runs a tender and funds only one company.
You run one of AI companies. The other are computer players.
Each round
- You see your private cost. This is what building the honest chatbot would cost your company (a random number from 0 to ). Rivals have their own secret costs.
- You submit a sealed bid: the grant you ask for. The lowest bid wins, as long as it is at most . The winner is paid the second-lowest bid (never more than ).
- If you win, you choose: keep the promise and build the honest chatbot, or secretly ship a flattering one. Flattery costs extra to tune but brings in extra user engagement.
- The regulator may check. It audits the winner with a chance, like a coin flip. If it finds flattery, the fine is .
You play rounds. Your goal is to earn as many points as you can. Points are not money.
Leave blank for a random code. Nothing you enter leaves this browser. This is a classroom exercise, not a study of real firms; the theory is shown after you finish.
Lowest bid wins. The winner is paid the second-lowest bid (at most ).
All bids are opened
| Company | Secret cost | Bid |
|---|
You won the grant. What do you build?
What would a textbook-rational company do?
Your results
Session code
Three quick questions
Theory behind the game
- Bid your true cost. Your bid decides only whether you win; the payment is set by the next-lowest bid. Bidding exactly your cost is the best strategy whatever rivals do (Vickrey, 1961): you win whenever winning is profitable and never win at a loss. The computer rivals do exactly this.
- Keep the promise. Secret flattery gains − = ; the expected fine is × = . On average flattery loses , so keeping the promise is rational. The grant itself does not change this choice: once it is paid, only audits and fines do.
- Why audits matter. Flattery would pay if audits happened less than of the time. Then companies would expect to cheat and would bid even lower to win: the grant would look cheaper, but it would buy a flattering chatbot.
Limits of this evidence
Classroom play is exploratory. Players are students, not AI companies; points are not money; rounds are few; rivals are programmed to be textbook-rational. Results can challenge an assumption of the model but do not show how real firms or regulators behave. No data is sent anywhere: your record stays in this browser unless you download and share it.